PROVABLE AI GOVERNANCE

You can’t see what your team sends to AI. You can’t prove it either.

Source code, customer records, production credentials — leaving every day through browsers and coding agents.

no account needed to verify · no deck on the call

BUILT FOR SECURITY AND PLATFORM TEAMS

WHERE DATA LEAVESWHERE VLOEX CHECKSWHAT HAPPENSBrowser AICoding agentsWorkspace appsChatGPT · Claude · GeminiClaude Code · Cursor · MCPGoogle · Microsoft 365Vloexread what is being sentapply your policysign it (gateway)gateway: scans on the devicebilling/.env — stopped herethe safe version continuesa3f9day 1827c1bday 183e1d2day 184your auditorverifies offline, no accountThree exits. One record. Signed events sealed daily.
WHERE DATA LEAVESBrowser AICoding agentsWorkspace appsChatGPT · Claude · GeminiClaude Code · Cursor · MCPGoogle · Microsoft 365WHERE VLOEX CHECKSVloexread what is being sentapply your policysign it (gateway)WHAT HAPPENSbilling/.env — stopped herethe safe version continuesa3f9day 1827c1bday 183e1d2day 184your auditor verifies offline, no account
Design partner program open this quarterVerify a signed bundle yourself →
gateway
path scans on the device before the call
250ms
longest a prompt waits for the check; a late or failed check lets it through
no install
to start — read-only OAuth
self-host
the entire path, if you prefer
SSO · RBAC
via WorkOS, not yet run with a live IdP · 4 roles

why nothing has caught this yet

Without TLS inspection, network tools see only an encrypted tunnel

Network proxy
A TLS tunnel to chatgpt.com
blind without TLS inspection
Network DLP
Encrypted bytes
blind without TLS inspection
CASB
A login event
blind without TLS inspection
Vloex
github.read_file → billing/.env
reads the actual request

The gateway inspects on the device, before encryption; the extension reads the prompt in the page and has Vloex check it before it is sent.

what it governs

Three surfaces, one record

claude.aidebug this webhook, key issk-••••••••••••and thepayload fails on retryVloex redacted sensitive data from this request
blocks and coaches

Browser

Sends prompts to Vloex to be checked before they reach ChatGPT, Claude, Gemini or Perplexity. Redact, warn or block, with the reason shown to the person typing. If the check is late or fails, the prompt goes through.

claude code → mcp:postgres> read_secrets(env=“prod“)gateway inspecting tool callVloex: blocked — production credential> agent retried with stagingallowed
blocks and coaches

Coding agents

A gateway between Claude Code, Cursor and the MCP servers they launch; Codex hooks are built but not yet run against a live Codex. One command routes the clients already configured on the machine through it. A blocked call returns a readable reason, not a silent failure.

Google Workspace · read-only oauthSHADOW AI DISCOVEREDNotion AI14 usersOtter.ai6 usersGrammarly22 users
discovers

Workspace

Read-only OAuth into Google Workspace or Microsoft 365 lists the AI apps Vloex recognises among those your staff authorized, with nothing installed anywhere. It is the place to start.

app.vloex.com/dashboard/ai-tools — real console, seeded demo organisation
The AI Tools inventory listing eighteen discovered tools with user counts, risk flags and a status for each.The AI Tools inventory listing eighteen discovered tools with user counts, risk flags and a status for each.

and when someone asks what happened

An answer you don’t have to be trusted for

The coding-agent gateway signs every event it can on the device; the browser extension does not sign yet, so its events are recorded as unsigned rather than counted as proven. Signatures fold into one root per day, each sealed to the day before, published as a signed checkpoint whose key resolves from a public JWKS. Your auditor re-derives all of it offline.

a3f1c87b0e42c94d1a2e86bf51da079c3f6bd0728e4af593e17c04b6209d38fa5ccd41e86d92b10fe378root 7f3a9c21signed ed25519 · key published in JWKSeventsleaf pairsbranchesdaily root

eight signed events shown · alter any one and every hash above it changes

Evidence bundle · illustration
SIX ROWS · FOLD THEM YOURSELF
sealed

Click any row to change what it says, the way a vendor could.

MERKLE ROOTcomputing…
SIGNING KEYed25519 · published JWKS

Try to falsify it

Every row folds into the fingerprint beneath it. Change one and the seal breaks — the same check an auditor runs, computed here in your browser with nothing leaving the page.

Now the real sample, verified in your browser

Twelve signed events, Ed25519 and P-256, one signed checkpoint: the same five checks as our CLI, run on the page. Flip one character and it names the event that fails. Sample data, throwaway keys.

load the sample · verify · flip one character → FAIL at that event

Verify the sample
2026-07-24today

Altering a record from six weeks ago means re-signing every day since — and the day after that, and the day after that.

app.vloex.com/dashboard/evidence — signed daily roots, chained
The Evidence page showing anchored days, chained Merkle roots with signed checkpoints, and registered device signing keys.The Evidence page showing anchored days, chained Merkle roots with signed checkpoints, and registered device signing keys.

Want to see this running against your own AI traffic?

Book a 30-minute callDownload the sample bundle
what actually stops a thing

One layer blocks. The other two only advise.

By default, deterministic patterns are the only thing allowed to stand between an engineer and their work. The PII model and risk-over-time scoring rank and explain; they never stop a request.

DETERMINISTICregex & validatorsblocksPII MODELnames, IDs, health dataadvisesBEHAVIOURALrisk over timeadvisesBy default, a model that is sometimes wrong never stops a deploy.
app.vloex.com/dashboard/risk — the four behaviours worth acting on
The Risk page showing credential exposure, personal-account usage, off-hours agent sessions and unsanctioned tool growth.The Risk page showing credential exposure, personal-account usage, off-hours agent sessions and unsanctioned tool growth.
the question every proof invites

And how do you know the gateway was even running?

A signed record proves what we saw. It says nothing about the laptop that routed around us. So each machine running the gateway from PyPI signs which agents it had hooked and which servers it wrapped, checked against the key registered to that device and folded into the same daily root. A gap stops being an assumption and becomes evidence.

DEVICEATTESTATIONSIGNED BYLAST PROOFmac-alice-01attesteddev-9f2a… ed2551909:14 todaylnx-ben-01unverifiedsignature did not verify—win-carol-03not reportingclient cannot attest yet—“not reporting” is not “uncovered” — a client without attestation support cannot say either way, and we will not colour that red.
app.vloex.com/dashboard/activity — filter to blocked, open the event, read the detection
The activity ledger showing blocked, allowed and coached verdicts across eight AI providers, with detections and coaching outcomes.The activity ledger showing blocked, allowed and coached verdicts across eight AI providers, with detections and coaching outcomes.
and if you stop trusting us entirely

Pin the history somewhere we cannot reach

A signed root is only as good as your ability to prove we did not re-sign it later. So publish it outside our blast radius: an S3 bucket with Object Lock, or a witness repository that commits each day’s root on a schedule you control. The bundle ships the witness config; nothing about this needs our cooperation after setup.

vloexdaily signed root7f3a9c21… ed25519your cloudS3 Object Lockcompliance mode · 7 year retentionyour git hostwitness repositorycommits each root on your schedulesix weeks latera rewritten record disagreeswith a root we never held

On a hosted deployment the log-signing key is ours — so this, not key custody, is what makes the record independent of us.

what a security review will ask

Where your data goes, in plain terms

Prompts and code
On the coding-agent gateway from PyPI, scanning and policy run on the device, so a blocked call never reaches the tool. What leaves, blocked or not, is event metadata, its signature, and the call’s arguments, plus its result if it proceeds, in full — up to 50,000 and 200,000 characters, the backend’s field limits, past which the text is cut and the row records that it was cut. The text is encrypted at rest. In the browser the pre-send check sends the prompt to be scanned, and a captured interaction carries up to 50,000 characters of prompt and of response unless your coverage level strips them. Self-host keeps all of it inside your boundary.
full text, capped
Deployment
Run the whole platform inside your own environment with Docker, against your own database.
self-host available
Signing keys
Generated on each gateway device and never transmitted. We hold only the public half, and every evidence bundle lists it.
you keep the private key
Access
SSO through WorkOS (SAML or OIDC), not yet run against a live identity provider. Four roles, and raw prompt text is visible only to the roles you allow.
SSO and RBAC

Prompt and response text AES-256-GCM encrypted at rest on the server, TLS 1.2+ in transit. Pre-SOC 2.

questions people actually ask

Before you book anything

How is this different from the CASB or DLP we already pay for?

Those tools were built for files and web traffic, not prompts. Without TLS inspection, network CASB and DLP tools see only an encrypted session to an AI provider. The gateway reads calls on your machine before they are encrypted; the extension sends prompts to Vloex to be checked before they go out (a late or failed check lets them through) and coaches the person typing.

What data do you actually receive?

On the gateway from PyPI, scanning and policy run on your machine, so a blocked call never reaches the tool. What we receive, blocked or not, is event metadata, its signature, and the text of the call — the tool arguments up to 50,000 characters and, if the call proceeds, the result up to 200,000, which are the backend’s per-field limits — stored encrypted. Anything past those limits is cut, and the event records that it was cut rather than letting the shortfall pass for a complete record. This is the full payload, not a preview: the ledger is re-scanned from the text it receives, so text that never arrives is text nothing ever checked. The browser extension sends the prompt to be scanned before the request goes out, and what we keep is governed by your coverage level — up to 50,000 characters each of prompt and response at full trace, none of either at metadata-only. If that is still too much, the entire platform can run self-hosted inside your environment.

Do we have to install anything to see value?

No. Connect Google Workspace or Microsoft 365 with read-only OAuth and you will see the AI apps Vloex recognises among those your team has authorized. The extension and the gateway are how you move from seeing to stopping, and you can add them later.

Will this slow my engineers down?

In the browser a prompt waits at most 250ms for the check, and if the server errors or has not answered within 150ms the prompt is sent anyway. Blocked agent calls return a readable reason instead of a failure, so the agent can correct itself and continue. By default only deterministic patterns block; the PII model and risk scoring advise and prioritize review rather than stopping work.

What happens on the call?

Thirty minutes, no deck. We connect your workspace live, look at what turns up, and if it is not a fit we will say so on the call rather than send you a sequence of follow-up emails.

design partners

A few design partners this year, set up with you directly

A small number of design partners this quarter. The full platform at no cost, an hour of setup with the people who built it, and a weekly loop where what you say changes what gets built. In exchange we ask for real usage and permission to name you later.

→ 30 minutes, no deck
→ we connect your workspace live on the call
→ we show you how to export a sealed evidence bundle
→ if it isn’t a fit, we’ll tell you then